Fix GPU unit test CI by disabling C++ tests and handling PG duplicate initialization (#2350)

Summary:
Pull Request resolved: #2350

More GPU unit test CI health fixes

Reviewed By: aporialiao

Differential Revision: D61984044

fbshipit-source-id: a5c590ae7aff5c92cf68e09655f0f94eeaaf66be

Author: PaulZhang12

.github/workflows/unittest_ci.yml

@@ -59,17 +59,3 @@ jobs:
           --ignore-glob='torchrec/inference/inference_legacy/tests*' --ignore-glob='*test_model_parallel_nccl*' \
           --ignore=torchrec/distributed/tests/test_cache_prefetch.py --ignore=torchrec/distributed/tests/test_fp_embeddingbag_single_rank.py \
           --ignore=torchrec/distributed/tests/test_infer_utils.py --ignore=torchrec/distributed/tests/test_fx_jit.py --ignore-glob=**/test_utils/
-        echo "Starting C++ Tests"
-        conda install -n build_binary -y gxx_linux-64
-        conda run -n build_binary \
-          x86_64-conda-linux-gnu-g++ --version
-        conda install -n build_binary -c anaconda redis -y
-        conda run -n build_binary redis-server --daemonize yes
-        mkdir cpp-build
-        cd cpp-build
-        conda run -n build_binary cmake \
-            -DBUILD_TEST=ON \
-            -DBUILD_REDIS_IO=ON \
-            -DCMAKE_PREFIX_PATH=/opt/conda/envs/build_binary/lib/python${{ matrix.python-version }}/site-packages/torch/share/cmake ..
-        conda run -n build_binary make -j
-        conda run -n build_binary ctest -V .

torchrec/test_utils/__init__.py

@@ -109,6 +109,8 @@ def init_distributed_single_host(
 ) -> dist.ProcessGroup:
     os.environ["LOCAL_WORLD_SIZE"] = str(local_size if local_size else world_size)
     os.environ["LOCAL_RANK"] = str(rank % local_size if local_size else rank)
+    if dist.is_initialized():
+        dist.destroy_process_group()
     dist.init_process_group(rank=rank, world_size=world_size, backend=backend)
     # pyre-fixme[7]: Expected `ProcessGroup` but got
     #  `Optional[_distributed_c10d.ProcessGroup]`.